Author: Robert Hyatt
Date: 07:21:11 01/29/04
Go up one level in this thread
On January 29, 2004 at 05:09:30, Tony Werten wrote: >On January 28, 2004 at 22:50:05, Robert Hyatt wrote: > >> >>Both are right. >> >>Technically, what happens is that it is possible for you to do a very deep >>search at the root, say when pondering, and you store a hash table entry for >>position "P" that says score >= .60, depth=15. Now after that long ponder, your >>opponent plays a different move. Each time you hit position P, you can use that >>>= .60 score because of the extreme draft stored in the table. And if your >>current beta value is (say) .3, then you will fail high since the table entry >>says >= .6. But when you relax beta, and re-search, now when you hit table >>entry P, you get sufficient draft, but the flag says "LOWER" which means that >>the .6 value stored is the LOWER bound. That is useless here since our UPPER >>bound (beta) is +infinity. You can't use it. And you are not searching deep >>enough to see the reason for the fail high, so now you fail low. >> >>That won't cause a problem if you implement it correctly, and the fail high _is_ >>the correct result for the best move. But you have to take care that the >>fail-low doesn't cause a re-search when you fail high again. And you have to be >>sure that you realize that after the fail-high, _that_ is the move you want to >>play even if it fails low on the re-search. > >Really ? I think I disagree. When this happens at the root you don't accept the >failhigh score, so why would you inside the tree ? Different animals. If I fail high on the aspiration window at the root, I _know_ that is a valid fail-high. And I _always_ accept that as the best move no matter _what_ happens on the re-search. If I fail high on the PVS null-window search at the root, it is common to immediately fail low on the re-search using the original aspiration window. I ignore that fail high completely as it is often false and caused by a null-move search failure somewhere below that node. I am not quite sure I understood your comment above, however, so maybe I missed your point "don't accept at root so why accept inside the tree?"..... > >Tony > >> >>Bruce's case is a pathological problem that will happen. But it is caused by an >>extreme happening. In a normal search this won't/can't happen (assuming you are >>not using null-move). But in reality it can. However, 99.9% of the time, >>re-searching with beta,+infinity after a fail high on alpha,beta will produce a >>score as expected...
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.